VM Context Switching Overhead

To measure context-switching overhead, we considered two workloads: a ``worst-case'' that cycles through a large memory buffer between switches, and a ``best-case'' that does not touch memory between switches. For the worst-case workload, context switch time starts at 3.9 $\mu$s for a single virtual machine, and increases to 9 $\mu$s for two or more VMs. For the best-case workload, the context switch time starts at 1.4 $\mu$s for a single virtual machine, and it increases slightly as the number of VMs increases; the slight increases coincide with the points at which the capacity of the L1 and L2 caches become exhausted. These results are commensurate with process context switching overheads in modern OSs.